References

• rather than objective, network-wide condition assessments. The South Sudan Infrastructure Development Authority (SIDA) has estimated that this informational deficit leads to suboptimal maintenance budget allocation causing approximately 25–30% excess lifecycle costs across the road network (SIDA, 2021). The development of a reliable, low-cost, and scalable road condition assessment methodology is therefore both a technical priority and a governance imperative.
• Several recent studies have demonstrated the viability of ML-based road condition assessment from satellite and aerial imagery in data-rich settings, including work by Maas and Rottensteiner (2016) using airborne LiDAR and multispectral data in Germany, Arya et al. (2021) applying deep learning to street- level imagery in India, and Owusu et al. (2021) using random forest classification on Sentinel-2 imagery in Ghana. However, applications specific to conflict-affected, data-scarce Sub-Saharan African environments — where ground truth data are sparse and the spectral complexity of degraded unpaved and gravel roads in tropical settings poses additional challenges — have not been reported in the literature.
• This paper makes the following contributions: (i) it develops and evaluates six ML models for four-class road condition classification using a 24-feature input derived from multispectral satellite imagery calibrated for South Sudan's road network; (ii) it applies a novel CNN+LSTM architecture that fuses spatial convolutional features with temporal sequence learning across six annual Sentinel-2 image composites (2019–2024) to exploit multi-temporal deterioration signals; (iii) it generates the first ML-predicted road condition map for the entire South Sudan classified network; and (iv) it quantifies temporal RCI deterioration trajectories for three strategic corridors, providing data directly applicable to maintenance programming.
• Road surface condition assessment from remote sensing has been pursued through several methodological approaches. Early work relied on thermal and multispectral aerial photography to detect surface distress features such as cracking and rutting (Brimley et al., 1996), but the spatial resolution limitations of early satellite platforms (30 m for Landsat TM) restricted applicability to detection of major surface failures rather than subtle condition gradations. With the advent of very-high-resolution commercial imagery (QuickBird, WorldView series, Pleiades) at 0.3–2.0 m resolution, several studies demonstrated the feasibility of automated crack and pothole detection using object-based image analysis (OBIA) and support vector machines (Radopoulou and Brilakis, 2016; Zhu et al., 2019). However, these approaches require expensive commercial imagery and are computationally intensive at national scales.
• The emergence of free, open, and regularly updated medium-resolution imagery — particularly Sentinel-2 (ESA, 2015) and Planet NICFI basemaps (Planet Labs, 2021) — has shifted the research frontier toward synoptic, network-scale assessment using less computationally demanding classifiers. Owusu et al. (2021) applied Random Forest to Sentinel-2 imagery in Ghana, achieving an overall accuracy of 79.4% for three-class road condition classification; Debella-Gilo and Etzelmüller (2022) used multi-temporal Sentinel-2 composite features to predict International Roughness Index (IRI) values for Kenyan national roads with a root mean square error of 1.8 m/km. Both studies highlight that spectral features alone are insufficient and that textural and contextual features significantly improve classification performance.
• Convolutional Neural Networks (CNNs) have become the dominant approach for image-based road condition assessment, exploiting their ability to learn hierarchical spatial feature representations directly from raw image data (LeCun et al., 1998). ResNet architectures, introduced by He et al. (2016), use residual skip connections to enable training of very deep networks (50–152 layers) without vanishing gradient problems, making them the backbone of choice for transfer learning from large annotated datasets (ImageNet, OpenStreetMap) to domain-specific applications with limited ground truth data.
• For temporal data, Long Short-Term Memory (LSTM) networks (Hochreiter and Schmidhuber, 1997) have shown particular efficacy in modelling sequential deterioration signals in time-series remote sensing data (Ienco et al., 2019). The combination of CNN spatial feature extraction with LSTM temporal modelling in hybrid CNN+LSTM architectures has been demonstrated to outperform either component alone for land cover change detection (Ji et al., 2019) and crop type mapping (Rußwurm and Körner, 2018), providing theoretical motivation for the architecture proposed in this paper.
• Road condition assessment in data-scarce, conflict-affected, and infrastructure-deficient settings poses distinctive challenges not present in well-studied environments. Ground truth data are sparse, potentially biased (surveys are more feasible on accessible roads, creating selection bias), and may be outdated relative to the most recent image acquisition. Spectral confusion between degraded unpaved road surfaces and surrounding bare soil or agricultural land is a significant problem in tropical environments where road-adjacent land use creates similar spectral signatures to road surfaces (Klonus et al., 2012). Semi-supervised and domain adaptation approaches have been proposed to address label scarcity (Tuia et al., 2016), but their application to road condition assessment in Sub-Saharan Africa remains largely unexplored. This study addresses these challenges through careful feature engineering, use of temporally stable contextual road buffer features, and a rigorous cross-validation strategy that explicitly accounts for spatial autocorrelation in the ground truth dataset.
• Sentinel-2 Level-2A (surface reflectance) imagery was acquired from the ESA Copernicus Open Access Hub for six dry-season annual composites (February–April, 2019–2024), using a least-cloud-pixel compositing approach within Google Earth Engine. Bands B2 (Blue, 10 m), B3 (Green, 10 m), B4 (Red, 10 m), B8 (NIR, 10 m), B11 (SWIR1, 20 m), and B12 (SWIR2, 20 m) were used, resampled to a common 10 m resolution. Planet NICFI high-resolution basemaps (4.77 m, RGB+NIR) for the same periods were acquired under the Nicfi Planet Data Programme for Tropical Forest Countries and used for texture feature extraction. A 30 m road buffer was applied to all road segments for feature extraction, excluding the first 5 m from the road centreline to reduce contamination by road-adjacent bare soil.
• Ground truth RCI values were obtained from the MoRB South Sudan Road Condition Survey 2022–23, which assessed 1,660 km of the classified network using the TRL Road Note 9 visual survey protocol, supplemented by rolling straightedge roughness measurements at 500 m intervals. RCI values were classified into four condition categories: Good (RCI 60–100), Fair (RCI 40–59), Poor (RCI 20–39), and Very Poor (RCI 0–19). The ground truth dataset was partitioned 70/15/15 into training, validation, and test sets, stratified by condition class and geographic region to minimise spatial autocorrelation bias. Table 1 summarises the class distribution in the ground truth dataset.
• For the CNN+LSTM model, the 24 per-image features were computed for each of the six annual composites (2019–2024), creating a temporal feature sequence of shape [6 × 24] per road segment. This temporal stack allows the LSTM component to learn deterioration trajectory patterns — for example, a segment showing progressive increase in BSI and GLCM Contrast over three consecutive years is more likely to be deteriorating than one showing stable values, even if the absolute value at the most recent time step is similar. Table 2 lists the complete feature set used for model training.
• Table 2: Complete input feature set (24 features per annual image composite) used for all machine learning models. For CNN+LSTM, temporal sequences of shape [6×24] are constructed across 2019–2024 annual composites.
• Random Forest (RF) was implemented using the scikit-learn library with 500 decision trees, maximum feature subset size of sqrt(24) = 5 features per split, and minimum samples per leaf of 5. Hyperparameters were tuned using a 5-fold stratified cross-validation grid search. XGBoost was implemented with 300 boosting rounds, learning rate η = 0.05, maximum tree depth of 6, and L1 regularisation parameter λ = 1.2. Both models used the 24 features from the most recent (2024) annual composite only, without the temporal stack used by CNN+LSTM.
• The CNN model was implemented as a ResNet-50 (He et al., 2016) adapted for 24-channel multispectral input (replacing the standard 3-channel RGB input). The network comprises an initial convolutional stem (7×7, 64 filters, stride 2), followed by four residual stages with 3, 4, 6, and 3 residual blocks respectively (64, 128, 256, 512 filters), global average pooling, and a fully connected classification head with softmax activation for four output classes. Input images were formed as 32×32 pixel patches centred on each 1 km road segment, tiled along the road buffer. Transfer learning from ImageNet weights was applied for the first three RGB channels, with random initialisation for the remaining 21 channels.
• The CNN+LSTM model processes the temporal feature sequence as follows. First, the ResNet-50 CNN encoder (weights shared across time steps) extracts a 512-dimensional spatial feature vector from each annual image patch. The resulting sequence of six feature vectors [h_2019, h_2020, ..., h_2024] is then passed to a two-layer bidirectional LSTM (128 hidden units per direction, dropout = 0.3) to model temporal dependencies. The final hidden state of the LSTM is concatenated with the most recent CNN feature vector and passed to a two-layer classification head (256 → 128 → 4 units) with ReLU activations and batch normalisation. The full model was trained end-to-end using the Adam optimiser (learning rate 5 × 10⁻⁴, weight decay 10⁻⁴) with a cosine annealing learning rate schedule over 80 epochs, and class-weighted cross-entropy loss to address class imbalance. The architectural forward pass for a single road segment is:
• h_t =CNN_encoder(X_t) for t = 2019, ..., 2024
• [c, h_T] = BiLSTM([h_2019, ..., h_2024])
• z = FC_head( CONCAT(h_T, h_2024) )
• The CNN+LSTM model was applied to all 8,314 road segments with available Sentinel-2 coverage, generating predicted RCI class labels and continuous RCI values for the full classified South Sudan road network. Results indicate that 64.1% of the network falls in the Poor or Very Poor condition category — a finding broadly consistent with, but slightly more pessimistic than, MoRB's own estimates based on the partial survey coverage (MoRB, 2022, reported 67% in poor/very poor). The predicted condition distribution is summarised in Table 4, disaggregated by road class.
• Table 4: Predicted Road Condition Distribution by Road Class — South Sudan Classified Network (2024)
• Table 4: Predicted road condition class distribution and weighted mean RCI for the South Sudan classified road network (8,400 km), disaggregated by road class. MoRB national target is a network-average RCI ≥ 40 by 2030.
• Figure 4 presents the temporal trajectory of predicted mean RCI for three priority corridors over the 2019–2024 study period, derived from applying the CNN+LSTM model to each annual image composite. All three corridors show consistent deterioration trends, with the N-8 Juba–Bor corridor declining from a predicted mean RCI of 42 in 2019 to 25 in 2024 — an average annual deterioration rate of approximately 3.4 RCI units per year. The Torit–Kapoeta corridor, while starting in better condition (predicted RCI = 55 in 2019), shows a slower but persistent deterioration rate of 2.2 units per year. These deterioration rates, combined with the current condition distribution in Table 4, enable projection of the condition under different maintenance funding scenarios.
• Figure 4: Left — Predicted vs. field-measured RCI for the CNN+LSTM model (test set, n=280 continuous RCI values), showing R²=0.924 and RMSE=5.8 RCI units. Right — Temporal mean RCI trends for three priority corridors, 2019–2024, derived from annual Sentinel-2 composite predictions.
• These deterioration rates, when extrapolated under the assumption of no maintenance intervention, project that the N-8 corridor will reach the Very Poor threshold (RCI = 20) by 2026 and the N-4 corridor by 2028. This quantitative deterioration analysis provides a directly actionable evidential basis for the prioritization of emergency rehabilitation investment on these corridors.
• The predicted finding that 64% of the South Sudan classified road network is in Poor or Very Poor condition has significant policy implications. If MoRB's 2030 target of a network-average RCI ≥ 40 is to be met, a minimum of 5,390 km requires rehabilitation or major maintenance within six years — approximately 900 km per year. At a conservative unit rehabilitation cost of USD 0.5 million/km for unsealed roads, this implies an annual rehabilitation budget requirement of USD 450 million, far exceeding current donor and government commitments of approximately USD 110 million/year. The predicted RCI map provides the spatial targeting information needed to direct available resources to the most critical segments, but the analysis also underscores the scale of the infrastructure deficit facing South Sudan.